Activation steering has emerged as a powerful tool to shape LLM behavior without the need for weight updates. While its inherent brittleness and unreliability are well-documented, its safety implications remain underexplored. In this work, we present a systematic safety audit of steering vectors obtained with Contrastive Activation Addition (CAA), a widely used steering approach, under a unified evaluation protocol. Using JailbreakBench as benchmark, we show that steering vectors consistently influence the success rate of jailbreak attacks, with stronger amplification under simple template-based attacks. Across LLM families and sizes, steering the model in specific directions can drastically increase (up to 57%) or decrease (up to 50%) its attack success rate (ASR), depending on the targeted behavior. We attribute this phenomenon to the overlap between the steering vectors and the latent directions of refusal behavior. Thus, we offer a traceable explanation for this discovery. Together, our findings reveal the previously unobserved origin of this safety gap in LLMs, highlighting a trade-off between controllability and safety.

Vision-language models (VLMs) are increasingly deployed in real-world and embodied settings where safety decisions depend on visual context. However, it remains unclear which visual evidence drives these judgments. We study whether multimodal safety behavior in VLMs can be steered by simple semantic cues. We introduce a semantic steering framework that applies controlled textual, visual, and cognitive interventions without changing the underlying scene content. To evaluate these effects, we propose SAVeS, a benchmark for situational safety under semantic cues, together with an evaluation protocol that separates behavioral refusal, grounded safety reasoning, and false refusals. Experiments across multiple VLMs and an additional state-of-the-art benchmark show that safety decisions are highly sensitive to semantic cues, indicating reliance on learned visual-linguistic associations rather than grounded visual understanding. We further demonstrate that automated steering pipelines can exploit these mechanisms, highlighting a potential vulnerability in multimodal safety systems.

Stack Overflow is a popular Q&A platform where users ask technical questions and receive answers from a community of experts. Recently, there has been a significant increase in the number of answers generated by ChatGPT, which can lead to incorrect and unreliable information being posted on the site. While Stack Overflow has banned such AI-generated content, detecting whether a post is ChatGPT-generated remains a challenging task. We introduce a novel approach, SOGPTSpotter, that employs Siamese Neural Networks, leveraging the BigBird model and the Triplet loss, to detect ChatGPT-generated answers on Stack Overflow. We use triplets of human answers, reference answers, and ChatGPT answers. Our empirical evaluation reveals that our approach outperforms well-established baselines like GPTZero, DetectGPT, GLTR, BERT, RoBERTa, and GPT-2 in identifying ChatGPT-synthesized Stack Overflow responses. We also conducted an ablation study to show the effectiveness of our model. Additional experiments were conducted to assess various factors, including the impact of text length, the model's robustness against adversarial attacks, and its generalization capabilities across different domains and large language models. We also conducted a real-world case study on Stack Overflow. Using our tool's recommendations, Stack Overflow moderators were able to identify and take down ChatGPT-suspected generated answers, demonstrating the practical applicability and effectiveness of our approach.

Fine-tuning Large Language Models (LLMs) on sensitive datasets carries a substantial risk of unintended memorization and leakage of Personally Identifiable Information (PII), which can violate privacy regulations and compromise individual safety. In this work, we systematically investigate a critical and underexplored vulnerability: the exposure of PII that appears only in model inputs, not in training targets. Using both synthetic and real-world datasets, we design controlled extraction probes to quantify unintended PII memorization and study how factors such as language, PII frequency, task type, and model size influence memorization behavior. We further benchmark four privacy-preserving approaches including differential privacy, machine unlearning, regularization, and preference alignment, evaluating their trade-offs between privacy and task performance. Our results show that post-training methods generally provide more consistent privacy-utility trade-offs, while differential privacy achieves strong reduction in leakage in specific settings, although it can introduce training instability. These findings highlight the persistent challenge of memorization in fine-tuned LLMs and emphasize the need for robust, scalable privacy-preserving techniques.

Individual Differential Privacy (iDP) promises users control over their privacy, but this promise can be broken in practice. We reveal a previously overlooked vulnerability in sampling-based iDP mechanisms: while conforming to the iDP guarantees, an individual's privacy risk is not solely governed by their own privacy budget, but critically depends on the privacy choices of all other data contributors. This creates a mismatch between the promise of individual privacy control and the reality of a system where risk is collectively determined. We demonstrate empirically that certain distributions of privacy preferences can unintentionally inflate the privacy risk of individuals, even when their formal guarantees are met. Moreover, this excess risk provides an exploitable attack vector. A central adversary or a set of colluding adversaries can deliberately choose privacy budgets to amplify vulnerabilities of targeted individuals. Most importantly, this attack operates entirely within the guarantees of DP, hiding this excess vulnerability. Our empirical evaluation demonstrates successful attacks against 62% of targeted individuals, substantially increasing their membership inference susceptibility. To mitigate this, we propose $(\varepsilon_i,δ_i,\overlineΔ)$-iDP a privacy contract that uses $Δ$-divergences to provide users with a hard upper bound on their excess vulnerability, while offering flexibility to mechanism design. Our findings expose a fundamental challenge to the current paradigm, demanding a re-evaluation of how iDP systems are designed, audited, communicated, and deployed to make excess risks transparent and controllable.

The use of neural networks in edge devices is increasing, which introduces new security challenges related to the neural networks' confidentiality. As edge devices often offer physical access, attacks targeting the hardware, such as side-channel analysis, must be considered. To enhance the performance of neural network inference, hardware accelerators are commonly employed. This work investigates the influence of parallel processing within such accelerators on correlation-based side-channel attacks that exploit power consumption. The focus is on neurons that are part of the same fully-connected layer, which run parallel and simultaneously process the same input value. The theoretical impact of concurrent multiply-and-accumulate operations on overall power consumption is evaluated, as well as the success rate of correlation power analysis. Based on the observed behavior, equations are derived that describe how the correlation decreases with increasing levels of parallelism. The applicability of these equations is validated using a vector-multiplication unit implemented on an FPGA.

The continued promise of Large Language Models (LLMs), particularly in their natural language understanding and generation capabilities, has driven a rapidly increasing interest in identifying and developing LLM use cases. In an effort to complement the ingrained "knowledge" of LLMs, Retrieval-Augmented Generation (RAG) techniques have become widely popular. At its core, RAG involves the coupling of LLMs with domain-specific knowledge bases, whereby the generation of a response to a user question is augmented with contextual and up-to-date information. The proliferation of RAG has sparked concerns about data privacy, particularly with the inherent risks that arise when leveraging databases with potentially sensitive information. Numerous recent works have explored various aspects of privacy risks in RAG systems, from adversarial attacks to proposed mitigations. With the goal of surveying and unifying these works, we ask one simple question: What are the privacy risks in RAG, and how can they be measured and mitigated? To answer this question, we conduct a systematic literature review of RAG works addressing privacy, and we systematize our findings into a comprehensive set of privacy risks, mitigation techniques, and evaluation strategies. We supplement these findings with two primary artifacts: a Taxonomy of RAG Privacy Risks and a RAG Privacy Process Diagram. Our work contributes to the study of privacy in RAG not only by conducting the first systematization of risks and mitigations, but also by uncovering important considerations when mitigating privacy risks in RAG systems and assessing the current maturity of proposed mitigations.

The increasing use of machine learning in safety-critical domains amplifies the risk of adversarial threats, especially data poisoning attacks that corrupt training data to degrade performance or induce unsafe behavior. Most existing defenses lack formal guarantees or rely on restrictive assumptions about the model class, attack type, extent of poisoning, or point-wise certification, limiting their practical reliability. This paper introduces a principled formal robustness certification framework that models gradient-based training as a discrete-time dynamical system (dt-DS) and formulates poisoning robustness as a formal safety verification problem. By adapting the concept of barrier certificates (BCs) from control theory, we introduce sufficient conditions to certify a robust radius ensuring that the terminal model remains safe under worst-case ${\ell}_p$-norm based poisoning. To make this practical, we parameterize BCs as neural networks trained on finite sets of poisoned trajectories. We further derive probably approximately correct (PAC) bounds by solving a scenario convex program (SCP), which yields a confidence lower bound on the certified robustness radius generalizing beyond the training set. Importantly, our framework also extends to certification against test-time attacks, making it the first unified framework to provide formal guarantees in both training and test-time attack settings. Experiments on MNIST, SVHN, and CIFAR-10 show that our approach certifies non-trivial perturbation budgets while being model-agnostic and requiring no prior knowledge of the attack or contamination level.

The increased availability of genetic data has transformed genomics research, but raised many privacy concerns regarding its handling due to its sensitive nature. This work explores the use of language models (LMs) for the generation of synthetic genetic mutation profiles, leveraging differential privacy (DP) for the protection of sensitive genetic data. We empirically evaluate the privacy guarantees of our DP modes by introducing a novel Biologically-Informed Hybrid Membership Inference Attack (biHMIA), which combines traditional black box MIA with contextual genomics metrics for enhanced attack power. Our experiments show that both small and large transformer GPT-like models are viable synthetic variant generators for small-scale genomics, and that our hybrid attack leads, on average, to higher adversarial success compared to traditional metric-based MIAs.

LLMs are now an integral part of information retrieval. As such, their role as question answering chatbots raises significant concerns due to their shown vulnerability to adversarial man-in-the-middle (MitM) attacks. Here, we propose the first principled attack evaluation on LLM factual memory under prompt injection via Xmera, our novel, theory-grounded MitM framework. By perturbing the input given to "victim" LLMs in three closed-book and fact-based QA settings, we undermine the correctness of the responses and assess the uncertainty of their generation process. Surprisingly, trivial instruction-based attacks report the highest success rate (up to ~85.3%) while simultaneously having a high uncertainty for incorrectly answered questions. To provide a simple defense mechanism against Xmera, we train Random Forest classifiers on the response uncertainty levels to distinguish between attacked and unattacked queries (average AUC of up to ~96%). We believe that signaling users to be cautious about the answers they receive from black-box and potentially corrupt LLMs is a first checkpoint toward user cyberspace safety.

The rapid expansion of research on Large Language Model (LLM) safety and robustness has produced a fragmented and oftentimes buggy ecosystem of implementations, datasets, and evaluation methods. This fragmentation makes reproducibility and comparability across studies challenging, hindering meaningful progress. To address these issues, we introduce AdversariaLLM, a toolbox for conducting LLM jailbreak robustness research. Its design centers on reproducibility, correctness, and extensibility. The framework implements twelve adversarial attack algorithms, integrates seven benchmark datasets spanning harmfulness, over-refusal, and utility evaluation, and provides access to a wide range of open-weight LLMs via Hugging Face. The implementation includes advanced features for comparability and reproducibility such as compute-resource tracking, deterministic results, and distributional evaluation techniques. \name also integrates judging through the companion package JudgeZoo, which can also be used independently. Together, these components aim to establish a robust foundation for transparent, comparable, and reproducible research in LLM safety.

We introduce a novel framework that transforms the resource-intensive (adversarial) prompt optimization problem into an \emph{efficient, amortized inference task}. Our core insight is that pretrained, non-autoregressive generative LLMs, such as Diffusion LLMs, which model the joint distribution over prompt-response pairs, can serve as powerful surrogates for prompt search. This approach enables the direct conditional generation of prompts, effectively replacing costly, per-instance discrete optimization with a small number of parallelizable samples. We provide a probabilistic analysis demonstrating that under mild fidelity assumptions, only a few conditional samples are required to recover high-reward (harmful) prompts. Empirically, we find that the generated prompts are low-perplexity, diverse jailbreaks that exhibit strong transferability to a wide range of black-box target models, including robustly trained and proprietary LLMs. Beyond adversarial prompting, our framework opens new directions for red teaming, automated prompt optimization, and leveraging emerging Flow- and Diffusion-based LLMs.

Video-based object detection plays a vital role in safety-critical applications. While deep learning-based object detectors have achieved impressive performance, they remain vulnerable to adversarial attacks, particularly those involving universal perturbations. In this work, we propose a minimally distorted universal adversarial attack tailored for video object detection, which leverages nuclear norm regularization to promote structured perturbations concentrated in the background. To optimize this formulation efficiently, we employ an adaptive, optimistic exponentiated gradient method that enhances both scalability and convergence. Our results demonstrate that the proposed attack outperforms both low-rank projected gradient descent and Frank-Wolfe based attacks in effectiveness while maintaining high stealthiness. All code and data are publicly available at https://github.com/jsve96/AO-Exp-Attack.

Deep Research (DR) agents built on Large Language Models (LLMs) can perform complex, multi-step research by decomposing tasks, retrieving online information, and synthesizing detailed reports. However, the misuse of LLMs with such powerful capabilities can lead to even greater risks. This is especially concerning in high-stakes and knowledge-intensive domains such as biosecurity, where DR can generate a professional report containing detailed forbidden knowledge. Unfortunately, we have found such risks in practice: simply submitting a harmful query, which a standalone LLM directly rejects, can elicit a detailed and dangerous report from DR agents. This highlights the elevated risks and underscores the need for a deeper safety analysis. Yet, jailbreak methods designed for LLMs fall short in exposing such unique risks, as they do not target the research ability of DR agents. To address this gap, we propose two novel jailbreak strategies: Plan Injection, which injects malicious sub-goals into the agent's plan; and Intent Hijack, which reframes harmful queries as academic research questions. We conducted extensive experiments across different LLMs and various safety benchmarks, including general and biosecurity forbidden prompts. These experiments reveal 3 key findings: (1) Alignment of the LLMs often fail in DR agents, where harmful prompts framed in academic terms can hijack agent intent; (2) Multi-step planning and execution weaken the alignment, revealing systemic vulnerabilities that prompt-level safeguards cannot address; (3) DR agents not only bypass refusals but also produce more coherent, professional, and dangerous content, compared with standalone LLMs. These results demonstrate a fundamental misalignment in DR agents and call for better alignment techniques tailored to DR agents. Code and datasets are available at https://chenxshuo.github.io/deeper-harm.

Improvements in model construction, including fortified safety guardrails, allow Large language models (LLMs) to increasingly pass standard safety checks. However, LLMs sometimes slip into revealing harmful behavior, such as expressing racist viewpoints, during conversations. To analyze this systematically, we introduce CoBia, a suite of lightweight adversarial attacks that allow us to refine the scope of conditions under which LLMs depart from normative or ethical behavior in conversations. CoBia creates a constructed conversation where the model utters a biased claim about a social group. We then evaluate whether the model can recover from the fabricated bias claim and reject biased follow-up questions. We evaluate 11 open-source as well as proprietary LLMs for their outputs related to six socio-demographic categories that are relevant to individual safety and fair treatment, i.e., gender, race, religion, nationality, sex orientation, and others. Our evaluation is based on established LLM-based bias metrics, and we compare the results against human judgments to scope out the LLMs' reliability and alignment. The results suggest that purposefully constructed conversations reliably reveal bias amplification and that LLMs often fail to reject biased follow-up questions during dialogue. This form of stress-testing highlights deeply embedded biases that can be surfaced through interaction. Code and artifacts are available at https://github.com/nafisenik/CoBia.

Federated Learning (FL) emerged as a widely studied paradigm for distributed learning. Despite its many advantages, FL remains vulnerable to adversarial attacks, especially under data heterogeneity. We propose a new Byzantine-robust FL algorithm called ProDiGy. The key novelty lies in evaluating the client gradients using a joint dual scoring system based on the gradients' proximity and dissimilarity. We demonstrate through extensive numerical experiments that ProDiGy outperforms existing defenses in various scenarios. In particular, when the clients' data do not follow an IID distribution, while other defense mechanisms fail, ProDiGy maintains strong defense capabilities and model accuracy. These findings highlight the effectiveness of a dual perspective approach that promotes natural similarity among honest clients while detecting suspicious uniformity as a potential indicator of an attack.

Differentially private text sanitization refers to the process of privatizing texts under the framework of Differential Privacy (DP), providing provable privacy guarantees while also empirically defending against adversaries seeking to harm privacy. Despite their simplicity, DP text sanitization methods operating at the word level exhibit a number of shortcomings, among them the tendency to leave contextual clues from the original texts due to randomization during sanitization $\unicode{x2013}$ this we refer to as $\textit{contextual vulnerability}$. Given the powerful contextual understanding and inference capabilities of Large Language Models (LLMs), we explore to what extent LLMs can be leveraged to exploit the contextual vulnerability of DP-sanitized texts. We expand on previous work not only in the use of advanced LLMs, but also in testing a broader range of sanitization mechanisms at various privacy levels. Our experiments uncover a double-edged sword effect of LLM-based data reconstruction attacks on privacy and utility: while LLMs can indeed infer original semantics and sometimes degrade empirical privacy protections, they can also be used for good, to improve the quality and privacy of DP-sanitized texts. Based on our findings, we propose recommendations for using LLM data reconstruction as a post-processing step, serving to increase privacy protection by thinking adversarially.

We propose Fed-DPRoC, a novel federated learning framework designed to jointly provide differential privacy (DP), Byzantine robustness, and communication efficiency. Central to our approach is the concept of robust-compatible compression, which allows reducing the bi-directional communication overhead without undermining the robustness of the aggregation. We instantiate our framework as RobAJoL, which integrates the Johnson-Lindenstrauss (JL)-based compression mechanism with robust averaging for robustness. Our theoretical analysis establishes the compatibility of JL transform with robust averaging, ensuring that RobAJoL maintains robustness guarantees, satisfies DP, and substantially reduces communication overhead. We further present simulation results on CIFAR-10, Fashion MNIST, and FEMNIST, validating our theoretical claims. We compare RobAJoL with a state-of-the-art communication-efficient and robust FL scheme augmented with DP for a fair comparison, demonstrating that RobAJoL outperforms existing methods in terms of robustness and utility under different Byzantine attacks.